Window selection for accurate music source separation using REPET

Separating music into vocal and non-vocal (back ground music) components is a challenging task, though it has a wide range of applications. The repeating rhythmic and beat property of the music, along with other accompaniments, has been leveraged in the `Repeated pattern extraction technique' (REPET), for separating a song into its vocal and non-vocal content. The REPET method involves measuring the repeating period of the music, deriving a model using it, and creating a time-frequency mask for separating the background (music) from the original mixture, in order to finally obtain the non-repeating vocal (source) content. Selecting an appropriate windowing function is critical for achieving the high quality of music-source separation in this method. In this paper, we examine five different windowing functions such as rectangular, flat-top, Blackman and Hanning window along with the Hamming window that was used originally in REPET method. Experiments are conducted on different types of music excerpts, using these five different windowing functions in the REPET method. Performance evaluation of the quality of music-source separation is carried out using `Analysis of Variation' (ANOVA) of `Signal to Interference Ratio' (SIR). The results indicate that no one particular window can be considered as completely reliable, for the accurate separation of a music mixture. However, Hamming window for extracting the background content, and Blackman window for extracting the foreground (vocal) content, give relatively better results.